电力电子装置故障波形相似性度量的小波矩阵变换法

Wavelet-matrix transforming method for similarity measurement of fault waveform of electronic power devices

中文关键词  小波变换  奇异值分解  内积变换  相似时序序列  电力电子装置

英文关键词  wavelet transform  singular-value-decomposition  inner-product transform  time series similarity

中文摘要

      提出一种基于小波矩阵变换的时序序列相似度量方法, 并对该方法应用于电力电子装置故障波形相似性度量进行了抗噪性、灵敏度及相似值准确性分析. 方法首先采用小波变换将时序序列压缩到小波子空间, 再由K-L变换(Karhunen-Loeve transformation)提取样本时序序列的特征向量和正交基, 然后将分析时序序列通过内积变换映射到正交基中得到分析特征向量, 最后计算两个特征向量之间的欧式距离以判定时序序列的相似度. 以电力电子装置故障波形的相似度量为例, 实验表明该方法特征向量维数低, 抗噪性好于直接小波法30倍, 灵敏度是直接小波法1/3, 相似值准确性好于小波奇异值法. 该方法对于大规模时序序列的相似匹配和检索具有潜在的应用价值.

英文摘要

      Based on the wavelet and the matrix transformation, we propose a method for measuring the time series similarity for application in the fault waveform similarity of electronic power devices. The noise-rejection ability, the sensitivity and the accuracy of this method are discussed. By using the wavelet transformation, we compress the timeseries sequence into the wavelet subspace. The sample’s feature vector and the orthogonal basis of the sampled time-series sequence are obtained by K-L transformation(Karhunen-Loeve transformation). By taking the inner-product, the analyzed time-series sequence is projected into the orthogonal basis, and the analyzed feature vector is thus obtained. Finally, the similarity value is calculated by the Euclid distance between the sample’s feature vector and the analyzed feature vector. In the measurement of the similarity between the fault waveforms in electronic power devices, the experimental results show that the dimension of feature vectors is low by the proposed method. In addition, the noise-rejection ability of the proposed method is 30 times higher than that of the plain wavelet method, the sensitivity of the proposed method is 1/3 of that of the plain wavelet method, and the accuracy of similarity value of the proposed method is higher than that of the wavelet singular-value-decomposition method. The proposed method has potential value in similarity matching and indexing for lager time-series databases.